GPU versions of AlphaFold are now available on Peregrine. You can find the available versions using module avail AlphaFold, and you can load the latest version using module load AlphaFold/2.3.1-foss-2022a-CUDA-11.7.0.
The module provides a simple alphafold symlink that points to the run_alphafold.py script, which means you can simply run alphafold with all required options (run alphafold --help to get more information).
Note that the run_alphafold.py was tweaked a little bit, so that it knows where to find required commands like hhblits, hhsearch, jackhmmer, kalign. This means that you do not have to provide the paths to these executables with options like --hhblits_binary_path.
By default, AlphaFold will try to use a GPU, and it even fails on nodes without a GPU. In order to instruct AlphaFold to run without a GPU, add the following to your job script:
export OPENMM_RELAX=CPU
The module allows you to control the number of cores used by the hhblits (default: 4 cores) and jackhmmer (default: 8 cores) tools by setting the environment variables $ALPHAFOLD_HHBLITS_N_CPU and/or $ALPHAFOLD_JACKHMMER_N_CPU. You can override the default number of cores using, for instance, export ALPHAFOLD_HHBLITS_N_CPU=8. Do note that these tools seem to run slower on more than 4/8 cores, but this may depend on your workload.
The large database files for the different AlphaFold versions are available in version-specific subdirectories at /scratch/public/AlphaFold/.
If you want to use different databases, you can override the default data directory by using export ALPHAFOLD_DATA_DIR=/path/to/data.
Given the fact that the initialization phase of AlphaFold is very I/O intensive while the database files are being read, reading the files from the /scratch file system directly is very time-consuming. In order to alleviate this issue the database files have been stored in a smaller Zstandard (zstd) compressed SquashFS file system image. Using this image instead of the files on /data directly is faster. These database images (which are also specific to the version of AlphaFold that you want to use) can be found at:
/scratch/public/AlphaFold/2.3.1.zstd.sqsh
The image can be mounted to a given directory using the squashfuse tool, for which a module is loaded that should give slightly better performance:
mkdir $TMPDIR/alphafold_data squashfuse /scratch/public/AlphaFold/2.3.1.zstd.sqsh $TMPDIR/alphafold_data
Now the AlphaFold databases are accessible at $TMPDIR/alphafold_data. The image can be unmounted using:
fusermount -u $TMPDIR/alphafold_data
The I/O performance can be increased even further by copying the squashfs image file to fast local node storage first. All nodes have at least 1 TB of fast solid state storage available.
The local disk can be reached using the environment variable $TMPDIR within the job. And copying can be done using the command:
cp /scratch/public/AlphaFold/2.3.1.zstd.sqsh $TMPDIR
The directory will be automatically removed when the job has finished. The mount command then looks as follows:
mkdir $TMPDIR/alphafold_data squashfuse $TMPDIR/2.3.1.zstd.sqsh $TMPDIR/alphafold_data
The following minimal examples can be used to submit an AlphaFold job to a regular (CPU) node or a V100 GPU node.
#!/bin/bash #SBATCH --job-name=alphafold #SBATCH --time=04:00:00 #SBATCH --partition=regular #SBATCH --nodes=1 #SBATCH --cpus-per-task=8 #SBATCH --mem=16GB # Clean the module environment and load the squashfuse and AlphaFold module module purge module load AlphaFold/2.3.1-foss-2022a-CUDA-11.7.0 # Uncomment the following line(s) if you want to use different values for the number of cores used by hhblits/jackhmmer #export ALPHAFOLD_HHBLITS_N_CPU=8 # default: 4 #export ALPHAFOLD_JACKHMMER_N_CPU=4 # default: 8 # Use the CPU instead of a GPU export OPENMM_RELAX=CPU # Copy the squashfs image to $TMPDIR cp /scratch/public/AlphaFold/2.3.1.zstd.sqsh $TMPDIR # Create a mountpoint for the AlphaFold database in squashfs format mkdir $TMPDIR/alphafold_data # Mount the AlphaFold database squashfs image squashfuse $TMPDIR/2.3.1.zstd.sqsh $TMPDIR/alphafold_data # Set the path to the AlphaFold database export ALPHAFOLD_DATA_DIR=$TMPDIR/alphafold_data # Run AlphaFold alphafold --fasta_paths=query.fasta --max_template_date=2020-05-14 --output_dir=output # Unmount the database image fusermount -u $TMPDIR/alphafold_data
#!/bin/bash #SBATCH --job-name=alphafold #SBATCH --time=04:00:00 #SBATCH --partition=gpu #SBATCH --nodes=1 #SBATCH --cpus-per-task=12 #SBATCH --mem=120GB #SBATCH --gres=gpu:1 module purge module load AlphaFold/2.3.1-foss-2022a-CUDA-11.7.0 # Uncomment the following line(s) if you want to use different values for the number of cores used by hhblits/jackhmmer #export ALPHAFOLD_HHBLITS_N_CPU=8 # default: 4 #export ALPHAFOLD_JACKHMMER_N_CPU=4 # default: 8 # Uncomment the following line if you are not running on a GPU node #export OPENMM_RELAX=CPU # Copy the squashfs image with the AlphaFold database to fast local storage cp /scratch/public/AlphaFold/2.3.1.zstd.sqsh $TMPDIR # Create a mountpoint for the AlphaFold database in squashfs format mkdir $TMPDIR/alphafold_data # Mount the AlphaFold database squashfs image squashfuse $TMPDIR/2.3.1.zstd.sqsh $TMPDIR/alphafold_data # Set the path to the AlphaFold database export ALPHAFOLD_DATA_DIR=$TMPDIR/alphafold_data # Run AlphaFold alphafold --fasta_paths=query.fasta --max_template_date=2020-05-14 --output_dir=output # Unmount the database image fusermount -u $TMPDIR/alphafold_data